To make GPT Image 2.5 change only what you want changed, write two sentences: the one thing that should change, then everything that must stay exactly as it is. OpenAI describes the model as “better at editing only what you’ve asked for, while keeping the rest of the details the same”, and naming both halves is what makes that repeatable.
Below: the two routes, the prompt shape, the selection limit, the edit endpoint and two worked examples.
The two ways to start an edit
The help centre documents two entry points.
You can edit an image in two ways:
- Select part of the image with the selection tool, then describe the changes in chat.
- Skip the selection tool and describe the edit directly in the conversation panel.
A selection suits a located fix; a description suits a change you can name in words. Comments are the same instinct with a pin — see GPT Image 2.5 comment editing.
One rule covers both routes: “If you want the edit to apply to a specific area, include that area in your prompt.” A highlight is a hint; the words name the object.
- Keep an approved copy of the source image
- Name the target area in words as well as on screen
- Keep one change per round
Write the change and the preservation list
A prompt that only says what to change leaves the model free everywhere else. The shape that keeps an edit small has three parts:
Change: <the single element to replace or adjust>
Keep: <what must stay identical: subject, pose, lighting, camera angle, background, crop, colours>
Do not add: <anything that must not appear: new objects, text, people, style shifts>Three details make the template work.
- One change per round, so a failed result tells you which instruction failed
- The keep list states facts, not preferences: the same lens, the same light, the same crop
- Visible text is quoted character by character, because a plausible render can still spell it wrong
The developer version of the same idea is updating one element “while preserving the subject, composition, and brand treatment around it.” The reusable prompt patterns behind the template are in GPT Image 2.5 prompting basics.
What the selection tool does and does not do
The selection tool points; it does not protect. It says where the instruction applies, and it does not freeze the pixels around that area. The help centre states the limit in one sentence:
Highlights are not always precise, and edits may extend beyond the area you selected.
That sentence is why this guide keeps asking for a written keep list. In practice:
- Elements near the edit can shift with it: hands, shadows, reflections, text
- A highlight is not a layer mask, so an edit can land partly outside it
Treat a selection as a strong hint rather than a lock.
The same job through the API
The API does the same work with fields instead of controls. The edit endpoint takes the reference images, the instruction and an optional mask:
images # reference image(s) for the edit — required
prompt # the change plus the preservation list — required
mask # guides where the edit applies; not a pixel-level lock
input_fidelity # high / low — how strongly the input image is favoured
background # auto / opaque / transparent
model # gpt-image-2.5-flare or gpt-image-2.5-sunburst
moderation # low / auto
n # number of images to return
output_compression # compression level for the returned file
output_format # png / webp / jpeg
partial_images # streamed previews while the image is built
quality # quality tier
size # output dimensions, or auto
stream # stream partial images as they are produced
user # end-user identifier for abuse trackingTwo notes. mask guides the edit rather than locking it, so the change can extend beyond it — which is why prompt still carries the preservation list. Parameter names and accepted values change between versions; the official API reference is the source of truth.
Endpoint fields are in the GPT Image 2.5 API overview; the two models are compared in Sunburst vs Flare.
Two worked examples: outfit and seasons
The first example protects an identity; the second protects a building.
Example one: new outfit, same face
Start from an approved portrait, and put the face into the keep list first.
Change: replace the grey hoodie with a structured navy wool coat, same fit across the shoulders
Keep: the same face, hair and skin tone; the same pose, camera angle, focal length, light direction, background and crop
Do not add: no new accessories, no text, no logo, and no change to the backgroundCheck the face and the collar line: if the jaw softens, the edit moved off the garment. Read GPT Image 2.5 character consistency if the identity slips.
Example two: four seasons on one castle
Here the subject never changes and only the weather does. Fix the camera once, then run one round per season.
Round 1 — Change: summer light, clear sky, warm green foliage
Round 2 — Change: autumn light, low sun, orange and rust foliage
Round 3 — Change: winter light, bare branches, thin snow on the roofs
Round 4 — Change: spring light, blossom on the trees, soft haze
Keep in every round: the same castle structure, tower count, window layout, camera position, focal length, framing and horizon line
Do not add: no new buildings, no people, no text, no change to the compositionAnchoring the building and the framing is what makes four files read as one location. Check that the towers and the horizon sit in the same place in every frame.
When an edit drifts: a fix order
Drift looks like small changes you never asked for: a different crop, a moved shadow, a rewritten background. Work through this order.
- Go back to the last good version and start again from it
- Cut the request to one variable and re-run it
- Rewrite the keep list as facts: camera, light, geometry, colours
- Only then consider another model or a higher quality tier
The order follows what OpenAI describes for long conversations: “each new edit builds on the work you’ve already done without degrading image quality over time.” Building on a good frame is the mechanism.
Frequently asked questions
Can I edit only part of an image?
In practice, yes. Name the part in the prompt, point at it if that helps, and keep an approved copy of the frame.
Does a selection lock everything else?
No. A highlight directs the request to an area rather than protecting the pixels around it, and edits may extend beyond the selection.
How do I stop an edit from drifting?
Keep one change per round, restate the keep list, and return to the last good version when a result drifts.
Can I use image editing through the API?
Yes. The edit endpoint takes reference images, a prompt and an optional mask, and the mask guides rather than locks the edit.
Targeted editing in GPT Image 2.5 is a habit rather than a control: one change, one keep list, one comparison against the approved copy.
Point at the area if that helps, describe both halves every time, and keep the frame you already like.